Information processing device, information processing method, and recording medium

Abstract

The present invention provides an information processing device comprising an acquisition means that acquires data, a classification means that classifies the data into any of a plurality of classification candidate classes or a background class, a determination means that determines whether or not the data is data outside a prescribed distribution on the basis of a background likelihood ratio of the likelihood that the data belongs to a classification candidate class to the likelihood that the data belongs to the background class, and an output means that outputs information about the determination.

Description

Information processing device, information processing method, and recording medium

[0001] The present disclosure relates to the technical fields of an information processing device, an information processing method, and a recording medium.

[0002] As an apparatus for classifying data, an apparatus that classifies data using a likelihood ratio that the data belongs to a class is known. For example, Patent Document 1 discloses an information processing apparatus including: an acquisition unit that sequentially acquires multiple elements included in sequence data; a first calculation unit that calculates an index indicating to which of multiple classes each of the multiple elements likely belongs, taking into account two or more of the multiple elements; a second calculation unit that integrates the indexes of the multiple elements to calculate an integrated index indicating to which of the multiple classes each of the multiple elements likely belongs; and a classification unit that classifies the sequence data into one of the classes based on the integrated index.

[0003] International Publication No. 2020 / 194497

[0004] An object of this disclosure is to provide an information processing device, an information processing method, and a recording medium that aim to improve upon the techniques disclosed in prior art documents.

[0005] One aspect of an information processing device includes an acquisition means for acquiring data, a classification means for classifying the data into one of a plurality of classification candidate classes and a background class, a determination means for determining whether the data is outside a predetermined distribution based on a background likelihood ratio between the likelihood that the data belongs to the classification candidate class and the likelihood that the data belongs to the background class, and an output means for outputting information related to the determination.

[0006] One aspect of the information processing method is an information processing method for determining data used by an information processing device that includes an acquisition means for acquiring data and a classification means for classifying the data into one of a plurality of classification candidate classes and a background class, and determines whether the data is outside a predetermined distribution based on a background likelihood ratio between the likelihood that the data belongs to the classification candidate class and the likelihood that the data belongs to the background class, and outputs information related to the determination.

[0007] One aspect of the recording medium is an information processing method for determining data used by an information processing device having an acquisition means for acquiring data and a classification means for classifying the data into one of a plurality of classification candidate classes and a background class, in which a computer program is recorded to cause a computer to execute the information processing method, which determines whether the data is outside a predetermined distribution based on a background likelihood ratio between the likelihood that the data belongs to the classification candidate class and the likelihood that the data belongs to the background class, and outputs information related to the determination.

[0008] FIG. 1 is a block diagram showing an example of the configuration of an information processing device according to the present disclosure. FIG. 2 is a block diagram showing an example of the configuration of an information processing device according to the present disclosure. FIG. 3 is a flowchart showing an example of the information processing operation of an information processing device according to the present disclosure. FIG. 4 is a conceptual diagram showing an example of the information processing operation of an information processing device according to the present disclosure. FIG. 5 is a block diagram showing an example of the configuration of an information processing device according to the present disclosure. FIG. 6 is a flowchart showing an example of the information processing operation of an information processing device according to the present disclosure.

[0009] Hereinafter, embodiments of an information processing device, an information processing method, and a recording medium will be described with reference to the drawings. [1: First Embodiment]

[0010] A first embodiment of an information processing device, an information processing method, and a recording medium will be described below. Hereinafter, the first embodiment of the information processing device, the information processing method, and the recording medium will be described using an information processing device 1 according to this disclosure. [1-1: Configuration of Information Processing Device 1]

[0011] 1 is a block diagram showing the configuration of an information processing device 1 according to this disclosure. As shown in FIG. 1, the information processing device 1 includes an acquisition unit 11, a classification unit 12, a determination unit 13, and an output unit 14.

[0012] The acquisition unit 11 acquires data. The classification unit 12 classifies the data into one of a plurality of classification candidate classes and a background class. The classification candidate classes are classes that can be classified by learning using data that belongs to the classification candidate classes (referred to as "classification candidate data"). The background classes are classes that can be classified by learning using data prepared based on the classification candidate data (referred to as "background data").

[0013] In this embodiment, data including classification candidate data and background data is referred to as training data. Furthermore, the distribution of indices relating to the characteristics of the training data in a predetermined space is referred to as a "predetermined distribution." The indices relating to the characteristics of the training data may be feature quantities extracted from the training data. In this case, the distribution of the feature quantities of the training data is referred to as a "predetermined distribution."

[0014] Data corresponding to a domain defined by a given distribution is referred to as "in-distribution data." Thus, training data is in-distribution data. On the other hand, data corresponding to a domain different from the domain defined by the given distribution is referred to as "out-of-distribution (OOD) data."

[0015] The determination unit 13 determines whether the data is out of distribution based on the likelihood ratio between the likelihood that the data belongs to the classification candidate class and the likelihood that the data belongs to the background class (referred to as the "background likelihood ratio"). The output unit 14 outputs information related to the determination. [1-2: Technical Effects of the Information Processing Device 1]

[0016] The information processing device 1 according to the present disclosure determines whether data is out of distribution during data classification and outputs information related to the determination, thereby preventing erroneous classification. [2: Second Embodiment]

[0017] A second embodiment of an information processing device, an information processing method, and a recording medium will be described below. Hereinafter, the second embodiment of an information processing device, an information processing method, and a recording medium will be described using an information processing device 2 according to this disclosure. [2-1: Domain and Classifier Behavior]

[0018] A range defined by the distribution of features in a feature space is called a domain. For example, the distribution of features for data captured with different cameras may differ. Therefore, data captured with different cameras may correspond to different domains. Furthermore, the distribution of features for data observed at different times may differ. Therefore, data observed at different times may correspond to different domains. Furthermore, the distribution of features for handwritten characters written by different people may differ. Therefore, handwritten characters written by different people may correspond to different domains. Furthermore, the features of English spoken by native English speakers may differ from the features of English spoken by non-native English speakers. Therefore, English spoken by native English speakers and English spoken by non-native English speakers may correspond to different domains.

[0019] A classifier that performs class classification is constructed so that it can accurately classify data in a domain corresponding to the training data used to train the classifier (i.e., data within a distribution). In other words, the classifier is adjusted to adapt to data within a distribution. Therefore, data in a domain different from the domain corresponding to the training data (i.e., data outside the distribution) is unknown data to the classifier. Therefore, data outside the distribution is data that is difficult for the classifier to accurately classify. When data outside the distribution is input, the classification accuracy of the classifier often decreases. When data outside the distribution is input, the classifier is prone to making incorrect classifications. In other words, the operating accuracy of the classifier depends on the domain corresponding to the input data.

[0020] However, the data to be inferred may be out-of-distribution data. That is, data outside the distribution may be input to the classifier. A classifier constructed by deep learning may confidently classify data from a different domain even when input, resulting in an incorrect classification (overconfident problem). That is, when data outside the distribution is input to the classifier, the classification result is unreliable.

[0021] Thus, whether the data input to the classifier is in-distribution data or out-of-distribution data is important information. This disclosure provides a mechanism for detecting out-of-distribution data from data input to the classifier and reporting that the data is out-of-distribution data. [2-2: Introduction of background classes]

[0022] The following paper prepares a model trained on training data (normal model) and a model trained on data (background data) that has been manipulated from the training data, and discloses the behavior of each model depending on the input data. Data in the domain corresponding to the training data (within the distribution) and data outside the domain corresponding to the training data (outside the distribution) have different background statistical information. The background model outputs different likelihoods when data within the distribution is input and when data outside the distribution is input. The background model outputs a small likelihood when data outside the distribution is input. The following paper discloses that when the likelihood ratio between the likelihood output by the normal model and the likelihood output by the background model is calculated, the likelihood ratio becomes larger when data outside the distribution is input. [Reference] Likelihood Ratio for Out-of-Distribution Detection (NeurIPS2019) https: / / arxiv.org / abs / 1906.02845

[0023] In this embodiment, in consideration of the characteristics of the background model prepared in the above-mentioned document, a background class is introduced as a classification target of a classifier that performs classification based on a likelihood ratio. Specifically, when K classification candidate classes exist, a background class is added as the (K+1)th class, and classification is performed using a classifier that performs K+1 class classification. The classifier classifies data into either the K classification candidate classes or the background class.

[0024] Since the likelihood of the background class decreases when out-of-distribution data is input, the likelihood ratio between the background class and the classification candidate class (background likelihood ratio) increases when out-of-distribution data is input. In this embodiment, the input of out-of-distribution data is detected by utilizing the fact that the background likelihood ratio differs between when in-distribution data is input and when out-of-distribution data is input. [2-3: Configuration of Information Processing Device 2]

[0025] 2 is a block diagram showing the configuration of the information processing device 2. As shown in FIG. 2, the information processing device 2 includes a calculation device 21 and a storage device 22. The information processing device 2 may further include a communication device 23, an input device 24, and an output device 25. However, the information processing device 2 does not necessarily have to include at least one of the communication device 23, the input device 24, and the output device 25. The calculation device 21, the storage device 22, the communication device 23, the input device 24, and the output device 25 may be connected via a data bus 26.

[0026] The arithmetic device 21 includes, for example, at least one of a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), and an FPGA (Field Programmable Gate Array). The arithmetic device 21 reads a computer program. For example, the arithmetic device 21 may read a computer program stored in the storage device 22. For example, the arithmetic device 21 may read a computer program stored in a computer-readable, non-transitory recording medium using a recording medium reading device (e.g., an input device 24 described later) not shown in the drawings that is included in the information processing device 2. The arithmetic device 21 may acquire (i.e., download or read) the computer program from a device (not shown) located outside the information processing device 2 via the communication device 23 (or another communication device). The arithmetic device 21 executes the read computer program. As a result, logical functional blocks for executing operations to be performed by the information processing device 2 are realized within the arithmetic device 21. In other words, the arithmetic device 21 can function as a controller for realizing logical functional blocks for executing operations (in other words, processes) to be performed by the information processing device 2. The arithmetic device 21 may output information to devices (not shown), such as other computers or cloud servers, that are provided outside the information processing device 2, via the communication device 23 (or other communication devices).

[0027] FIG. 2 shows an example of logical functional blocks implemented within the arithmetic device 21 to perform information processing operations. As shown in FIG. 2 , the arithmetic device 21 includes an acquisition unit 211, which is a specific example of an "acquisition means" described in the appendix, a classification unit 212, which is a specific example of a "classification means" described in the appendix, a determination unit 213, which is a specific example of a "determination means" described in the appendix, an output unit 214, which is a specific example of an "output means" described in the appendix, and a calculation unit 215, which is a specific example of a "calculation means" described in the appendix. However, the calculation unit 215 does not necessarily need to be implemented within the arithmetic device 21. The classifier described above may be implemented by the classification unit 212 and the calculation unit 215. Learning of the classifier will be described in another embodiment. Details of the operations of the acquisition unit 211, classification unit 212, determination unit 213, output unit 214, and calculation unit 215 will be described later with reference to FIG. 3.

[0028] The storage device 22 can store desired data. For example, the storage device 22 may temporarily store a computer program executed by the arithmetic device 21. The storage device 22 may temporarily store data that the arithmetic device 21 temporarily uses when the arithmetic device 21 is executing a computer program. The storage device 22 may store data that the information processing device 2 stores long-term. The storage device 22 may include at least one of a RAM (Random Access Memory), a ROM (Read Only Memory), a hard disk device, a magneto-optical disk device, an SSD (Solid State Drive), and a disk array device. In other words, the storage device 22 may include a non-temporary recording medium.

[0029] The communication device 23 is capable of communicating with devices external to the information processing device 2 via a communication network (not shown). The communication device 23 may be a communication interface based on standards such as Ethernet (registered trademark), Wi-Fi (registered trademark), Bluetooth (registered trademark), or USB (Universal Serial Bus).

[0030] The input device 24 is a device that accepts information input to the information processing device 2 from outside the information processing device 2. For example, the input device 24 may include an operation device (e.g., at least one of a keyboard, a mouse, and a touch panel) that can be operated by an operator of the information processing device 2. For example, the input device 24 may include a reading device that can read information recorded as data on a recording medium that can be externally attached to the information processing device 2.

[0031] The output device 25 is a device that outputs information to the outside of the information processing device 2. For example, the output device 25 may output information as an image. That is, the output device 25 may include a display device (a so-called display) that can display an image showing the information to be output. For example, the output device 25 may output information as sound. That is, the output device 25 may include an audio device (a so-called speaker) that can output sound. For example, the output device 25 may output information on paper. That is, the output device 25 may include a printing device (a so-called printer) that can print desired information on paper. [2-4: Information Processing Operation Performed by the Information Processing Device 2]

[0032] The information processing operation performed by the information processing device 2 will be described with reference to Fig. 3. Fig. 3 is a flowchart showing an example of the flow of the information processing operation performed by the information processing device 2.

[0033] 3, the acquiring unit 211 acquires data (step S21). The data acquired in step S21 is data to be classified. The correct class of the data acquired in step S21 may be unknown. It may also be unknown whether the data acquired in step S21 is data within the distribution or data outside the distribution.

[0034] The data acquired in step S21 is input to the calculation unit 215. The calculation unit 215 calculates a likelihood ratio of the data (step S22). The likelihood ratio is a ratio between the likelihood that the data belongs to one of two classes out of a plurality of classification candidate classes and the background class and the likelihood that the data belongs to the other class. The likelihood ratio is an index indicating the likelihood that the data belongs to a certain class. The calculation unit 215 may calculate the logarithm of the likelihood ratio (log likelihood ratio (LLR)). Note that in the following description, the log likelihood ratio may be simply referred to as the likelihood ratio.

[0035] When there are three classification candidate classes, class 0, class 1, and class 2, the calculation unit 215 may calculate a likelihood ratio matrix as illustrated in Fig. 4. In the likelihood ratio matrix illustrated in Fig. 4, "0" represents the likelihood that data belongs to class 0, "1" represents the likelihood that data belongs to class 1, "2" represents the likelihood that data belongs to class 2, and "B" represents the likelihood that data belongs to the background class. Therefore, in the likelihood ratio matrix illustrated in Fig. 4, "1 / 0" represents the likelihood ratio between class 1 and class 0, "2 / 0" represents the likelihood ratio between class 2 and class 0, "B / 0" represents the likelihood ratio between the background class and class 0, "2 / 1" represents the likelihood ratio between class 2 and class 1, "B / 1" represents the likelihood ratio between the background class and class 1, and "B / 2" represents the likelihood ratio between the background class and class 2.

[0036] Among the likelihood ratios, the likelihood ratio between the likelihood that the data belongs to a candidate classification class and the likelihood that the data belongs to a background class (in the example shown in Figure 4, "B / 0", "B / 1", "B / 2", "0 / B", "1 / B", "2 / B") is referred to as the background likelihood ratio.

[0037] The determination unit 213 determines whether the data is out-of-distribution data based on the background likelihood ratio calculated by the calculation unit 215 (step S23). The determination unit 213 may compare the background likelihood ratio with a predetermined determination threshold to determine whether the data is out-of-distribution data. The determination unit 213 may determine that the data is out-of-distribution data when any one of the background likelihood ratios corresponding to the multiple classification candidate classes exceeds the predetermined determination threshold. For example, the determination unit 213 may determine that the data is out-of-distribution data when any one of the background likelihood ratios corresponding to the multiple classification candidate classes exceeds the predetermined determination threshold. Alternatively, the determination unit 213 may determine that the data is out-of-distribution data when all of the background likelihood ratios corresponding to the multiple classification candidate classes exceed the predetermined determination threshold. The timing at which the data is determined to be out-of-distribution data may be set according to the operating characteristics of the calculation unit 215 and the determination unit 213. For example, if the calculation unit 215 has been trained to easily calculate a relatively large background likelihood ratio, the data may be determined to be out-of-distribution data when the background likelihood ratios corresponding to all classification candidate classes exceed the determination threshold. Alternatively, if the calculation unit 215 has been trained to easily calculate a relatively large background likelihood ratio, the determination threshold may be set to a large value. The determination threshold may be a value common to multiple classification candidate classes. Alternatively, the determination threshold may be set for each of the multiple classification candidate classes.

[0038] If the data is out-of-distribution data (step S23: Yes), the output unit 214 outputs an alert as information related to the determination (step S24). The output unit 214 may control a display serving as the output device 25 to display an alert. The output unit 214 may control a speaker serving as the output device 25 to output an alert sound. The output unit 214 may control the storage device 22 to store information indicating that the acquired data is out-of-distribution data in the storage device 22. The output unit 214 may control the communication device 23 to transmit information indicating that the acquired data is out-of-distribution data to an external device.

[0039] If the data is not out-of-distribution data (step S23: No), the classification unit 212 classifies the data into one of a plurality of classification candidate classes and a background class based on the likelihood ratio calculated by the calculation unit 215 (step S25). The classification unit 212 may compare the likelihood ratio calculated by the calculation unit 215 with a predetermined classification threshold to classify the data into one of a plurality of classification candidate classes and a background class. The classification threshold may be the same value as the judgment threshold, or may be a different value. The classification threshold and the judgment threshold may be set according to the operating characteristics of the classification unit 212, the judgment unit 213, and the calculation unit 215. At least one of the judgment threshold and the classification threshold may be manually set. Furthermore, at least one of the judgment threshold and the classification threshold may be automatically and sequentially updated.

[0040] The output unit 214 outputs the classification result as information related to the determination (step S26). In step S26, the output unit 214 outputs information indicating the class to which the acquired data belongs. The output unit 214 may control a display serving as the output device 25 to display the class to which the acquired data belongs. The output unit 214 may control a speaker serving as the output device 25 to output the class to which the acquired data belongs by voice. The output unit 214 may control the storage device 22 to store the class to which the acquired data belongs in the storage device 22. The output unit 214 may control the communication device 23 to transmit the class to which the acquired data belongs to an external device. [2-5: Technical Effects of Information Processing Device 2]

[0041] The information processing device 2 according to this disclosure introduces a background class whose likelihood decreases when out-of-distribution data is input, and is therefore able to determine whether the input data is out-of-distribution data. The information processing device 2 can perform a class classification operation and a determination operation as to whether the data is out-of-distribution data based on a comparison of the likelihood ratios of a plurality of classification candidate classes and the background class with a predetermined threshold. The information processing device 2 can also perform a classification operation according to the characteristics of a calculation model and a determination operation according to the characteristics of the calculation model. [3: Third Embodiment]

[0042] A third embodiment of an information processing device, an information processing method, and a recording medium will be described. Hereinafter, a third embodiment of an information processing device, an information processing method, and a recording medium will be described using an information processing device 3 according to this disclosure.

[0043] As shown in FIG. 5 , the information processing device 3 differs from the information processing device 2 in that a learning unit 316 is further implemented within the calculation device 21. Other features of the information processing device 3 may be the same as other features of the information processing device 2. The learning unit 316 performs learning related to the classification operation by the classifier. The classifier may be implemented by the calculation unit 315 and the classification unit 312. The learning unit 316 performs at least one of learning related to the calculation of likelihood ratios by the calculation unit 315 and learning related to the classification of data by the classification unit 312. [3-1: Information Processing Operation Performed by the Information Processing Device 3]

[0044] The flow of the information processing operation performed by the information processing device 3 will be described with reference to Fig. 6. Fig. 6 is a flowchart showing the flow of the information processing operation performed by the information processing device 3.

[0045] As shown in FIG. 6 , the acquiring unit 311 acquires training data (step S31). The acquiring unit 311 acquires training data including sample data of either classification candidate data corresponding to a plurality of classification candidate classes or background data prepared using the classification candidate data. The background data includes at least one of data obtained by perturbing the classification candidate data and data obtained by randomly mixing the classification candidate data. The background data includes statistical information about the background of the domain corresponding to the training data. Learning background classes may also be rephrased as learning statistical information about the background according to the domain corresponding to the training data. The training data may be, for example, a dataset of sample data of either the plurality of classification candidate data or the background data, and information about the correct class to which the sample data belongs.

[0046] The learning unit 316 inputs sample data included in the training data to the calculation unit 315. The calculation unit 315 calculates a likelihood ratio of the input sample data (step S32). The calculation unit 315 may calculate the likelihood ratio of the input sample data using a calculation model.

[0047] The classification unit 312 classifies the sample data included in the training data based on the likelihood ratio calculated by the calculation unit 315 (step S33). The classification unit 312 classifies the sample data included in the training data into one of a plurality of classification candidate classes or a background class.

[0048] The learning unit 316 performs learning on the classification operation by the classifier (step S34). The learning unit 316 performs at least one of learning on the calculation of likelihood ratios by the calculation unit 315 and learning on data classification by the classification unit 312.

[0049] The learning unit 316 may perform learning on the classification operation of the classifier based on the correct class included in the training data and the estimated class of the sample data classified by the classifying unit 312. The learning unit 316 may calculate the error between the correct class and the estimated class using, for example, a cross-entropy loss function, and perform learning based on this error.

[0050] Based on the above-described error, the learning unit 316 may update the values ​​of the parameters included in the calculation model used by the calculation unit 315. For example, the learning unit 316 may optimize the values ​​of the parameters included in the calculation model used by the calculation unit 315 so that the value of the above-described error is minimized. Note that, as a parameter optimization method using a loss function, for example, backpropagation may be used, but other methods may also be used.

[0051] The learning unit 316 determines whether learning has ended (step S35). The learning unit 316 may determine whether learning has ended, for example, based on whether all learning data has been input. Alternatively, the learning unit 316 may determine whether learning has ended based on whether a predetermined period of time has elapsed since the start of learning. Alternatively, the learning unit 316 may determine whether learning has ended based on whether the processes of steps S31 to S34 described above have been looped a predetermined number of times.

[0052] If it is determined that the learning has been completed (step S35: YES), the series of processes ends. The learning unit 316 may store the parameters included in the calculation model used by the calculation unit 315 in the storage device 22.

[0053] On the other hand, if it is determined that the learning has not been completed (step S35: NO), the process returns to step S31. This allows the learning process using the learning data to be repeated, improving the accuracy of the classification operation by the classifier. [3-2: Technical Effects of the Information Processing Device 3]

[0054] The information processing device 3 according to this disclosure can construct a classifier that classifies data into one of a plurality of classification candidate classes or a background class by using classification candidate data corresponding to a plurality of classification candidate classes and background data prepared using the classification candidate data. The background data includes at least one of data obtained by perturbing the classification candidate data and data obtained by randomly mixing the classification candidate data, so that it is possible to learn a background class related to the characteristics of a domain corresponding to the training data. [4: Fourth Embodiment]

[0055] Next, a fourth embodiment of an information processing device, an information processing method, and a recording medium will be described. Hereinafter, the fourth embodiment of an information processing device, an information processing method, and a recording medium will be described using an information processing device 4 according to this disclosure.

[0056] 7 , the information processing device 4 further includes a repeating unit 417 in the calculation unit 21. The repeating unit 417 repeatedly executes the operations of the acquiring unit 411, the calculating unit 415, the determining unit 413, and the classifying unit 412 until an operation completion condition is satisfied. The operation completion condition is satisfied when at least one of the following occurs: the determining unit 413 determines that the sequence data is outside the predetermined distribution; and the classifying unit 412 confirms the classification of the sequence data.

[0057] In this embodiment, the data is sequence data. In this embodiment, sequence data is data that includes multiple elements arranged in a predetermined order, such as time-series data. More specific examples of sequence data include, but are not limited to, video data, audio data, or subdivided image data.

[0058] The information processing device 4 performs an operation using a sequential probability ratio test (SPRT). Specifically, at least one of the calculation of likelihood ratios by the calculation unit 415 and the classification of data by the classification unit 412 is performed using SPRT. The information processing device 4 may perform a classification operation using a sequential probability ratio test-based algorithm that treats as an Nth-order Markov series (SPRT-TANDEM), which treats a data series as an Nth-order Markov process. In other words, this embodiment may be said to introduce a background class as the (K+1)th class into SPRT-TANDEM, which performs K-class classification. [4-1: Inference Operation]

[0059] As shown in FIG. 8 , the acquisition unit 411 acquires elements included in the sequence data (step S41). The acquisition unit 411 is configured to sequentially acquire a finite number of elements included in the sequence data. For example, the acquisition unit 411 may be configured to acquire elements included in the sequence data one by one in order. The acquisition unit 411 may acquire data directly from any data acquisition device (e.g., a camera, a microphone, etc.), or may read data that has been previously acquired by a data acquisition device and stored in the storage device 22, etc. When the acquisition unit 411 acquires data from a camera, the acquisition unit 411 may be configured to acquire data from each of multiple cameras. The acquisition unit 411 outputs the multiple elements to the calculation unit 415.

[0060] The calculation unit 415 calculates a likelihood ratio based on at least two elements (step S42). The at least two elements include an element acquired in the current repetitive operation and an element processed in a previous repetitive operation. First, the calculation unit 415 calculates a first likelihood ratio based on the at least two elements. Next, the calculation unit 415 calculates a current second likelihood ratio by integrating the second likelihood ratio calculated in the previous repetitive operation and the first likelihood ratio calculated in the current repetitive operation. The likelihood ratio calculated in step S42 is the current second likelihood ratio.

[0061] For example, the element obtained in this iteration (x Ｎ ) and the element processed in the previous iteration (x Ｎ－１ When two elements (x) are taken into consideration, the second likelihood ratio can be calculated using the following formula 1. ｉ ) is Class C １ The likelihood of belonging to p(x ｉ |C １ ) and the element (x ｉ ) is Class C ２ The likelihood of belonging to p(x ｉ |C ２ ) [Equation 1]

[0062] In addition, the element obtained in this iteration is compared with the element processed in the previous iteration (x Ｎ－２ When the three factors up to (1) and (2) are taken into consideration, the second likelihood ratio can be calculated using the following formula 2. [Formula 2]

[0063] The terms on the right side of the above formulas 1 and 2 are the first likelihood ratios, and the terms on the left side of the above formulas 1 and 2 are the second likelihood ratios. １ and the likelihood of belonging to class C ２ In step S42, the calculation unit 415 may calculate a likelihood ratio matrix as shown in FIG.

[0064] The determination unit 413 compares the background likelihood ratio of the likelihood ratios calculated by the calculation unit 415 with a predetermined determination threshold to determine whether the data is out-of-distribution data (step S43). The determination unit 413 may determine that the data is out-of-distribution data when any one of the plurality of background likelihood ratios corresponding to the plurality of classification candidate classes exceeds the predetermined determination threshold. The determination unit 413 may determine that the data is out-of-distribution data when any one of the plurality of background likelihood ratios corresponding to the plurality of classification candidate classes exceeds the predetermined determination threshold. Alternatively, the determination unit 413 may determine that the data is out-of-distribution data when all of the plurality of background likelihood ratios corresponding to the plurality of classification candidate classes exceed the predetermined determination threshold.

[0065] If the data is out-of-distribution data (step S43: Yes), the output unit 214 outputs an alert as information related to the determination (step S44). If the data is not out-of-distribution data (step S43: No), the classification unit 412 classifies the data into one of the plurality of classification candidate classes or the background class based on the likelihood ratio calculated by the calculation unit 215.

[0066] The classification unit 412 determines whether the classification is complete (step S45). The classification unit 412 may determine whether the classification is complete by comparing the likelihood ratio calculated by the calculation unit 215 with a predetermined classification threshold.

[0067] If the classification is complete (step S45: Yes), the operations of the acquisition unit 411, calculation unit 415, determination unit 413, and classification unit 412 end. The classification may be complete when the classification unit 412 has confirmed the classification of the sequence data into one of the multiple classification candidate classes or the background class. If any likelihood ratio reaches a predetermined classification threshold, the classification unit 412 may confirm the classification of the sequence data into one of the multiple classification candidate classes or the background class. The output unit 214 outputs the classification result (step S46).

[0068] If the classification is not completed (step S45: No), the process returns to step S41, and the repeating unit 417 causes the acquiring unit 411, the calculating unit 415, the determining unit 413, and the classifying unit 412 to perform their operations. [4-2: Learning Operation]

[0069] The learning unit 416 performs learning using a loss function that takes into account likelihood ratios of (K+1)×K patterns, where the likelihood that the sequence data belongs to one of the K+1 classes is used as the denominator and the likelihood that the sequence data belongs to another class is used as the numerator. The learning unit 416 may also perform learning using a loss function that takes into account the likelihood ratio of the (K+1)×K patterns in which the correct class is used as the numerator.

[0070] The learning unit 416 may perform learning related to the calculation of the likelihood ratio using a loss function (LLLR) in which the likelihood ratio increases when the correct class to which the sequence data belongs is in the numerator of the likelihood ratio and decreases when the correct class is in the denominator of the likelihood ratio. The learning unit 416 may also perform learning related to the calculation of the likelihood ratio using a log-sum-exponential loss function (LSEL). That is, the learning unit 416 performs learning related to the calculation of the likelihood ratio using at least one of LLLR and LSEL. [4-3: Technical Effects of the Information Processing Device 4]

[0071] The information processing device 4 according to this disclosure uses SPRT, and therefore can quickly determine to which class the sequence data belongs or whether the sequence data is out-of-distribution data.

[0072] When LLLR is used for likelihood ratio training, training can be performed so that the penalty for an incorrect class becomes large and the penalty for a correct class becomes small. As a result, the information processing device 4 can appropriately select at least one class to which the sequence data belongs from among multiple classes that are classification candidates.

[0073] Furthermore, when LSEL is used for likelihood ratio learning, the convergence of the stochastic gradient descent method is improved. More specifically, a larger gradient can be assigned to items that are relatively difficult to classify using likelihood ratios, thereby accelerating convergence. Therefore, by using LSEL, the information processing device 4 can perform efficient learning. [5: Supplementary Note]

[0074] The following supplementary notes are further disclosed with respect to the above-described embodiments. [Supplementary Note 1] An information processing device comprising: an acquisition means for acquiring data; a classification means for classifying the data into one of a plurality of classification candidate classes and a background class; a determination means for determining whether the data is outside a predetermined distribution based on a background likelihood ratio between the likelihood that the data belongs to the classification candidate class and the likelihood that the data belongs to the background class; and an output means for outputting information related to the determination. [Supplementary Note 2] The information processing device according to Supplementary Note 1, further comprising: a calculation means for calculating a likelihood ratio between the likelihood that the data belongs to one of two classes among the plurality of classification candidate classes and the background class and the likelihood that the data belongs to the other class. [Supplementary Note 3] The information processing device according to Supplementary Note 2, wherein the classification means classifies the data based on the likelihood ratio calculated by the calculation means. [Supplementary Note 4] The information processing device according to Supplementary Note 2, wherein the determination means makes a determination based on the background likelihood ratio calculated by the calculation means. [Supplementary Note 5] The information processing device according to Supplementary Note 1, wherein the determination means compares the background likelihood ratio with a predetermined determination threshold to determine whether the data is outside the predetermined distribution. [Supplementary Note 6] The information processing device according to Supplementary Note 2, further comprising learning means that performs at least one of learning related to calculation of the likelihood ratio by the calculation means and learning related to classification of the data by the classification means. [Supplementary Note 7] The information processing device according to Supplementary Note 6, wherein the learning means performs learning using training data including classification candidate data corresponding to the plurality of classification candidate classes and background data prepared using the classification candidate data. [Supplementary Note 8] The information processing device according to Supplementary Note 7, wherein the background data includes at least one of data obtained by perturbing the classification candidate data and data obtained by randomly mixing the classification candidate data. [Supplementary Note 9] The information processing device according to Supplementary Note 7, wherein the predetermined distribution is defined by the distribution of features of the training data in feature space.[Supplementary Note 10] The information processing device according to Supplementary Note 1, wherein the determination means determines that the data is out of the predetermined distribution when any one of the plurality of background likelihood ratios corresponding to the plurality of classification candidate classes exceeds a predetermined determination threshold, and the output means outputs an alert when the data is determined to be out of the predetermined distribution. [Supplementary Note 11] The information processing device according to Supplementary Note 1, wherein the determination means determines that the data is out of the predetermined distribution when any one of the plurality of background likelihood ratios corresponding to the plurality of classification candidate classes exceeds a predetermined determination threshold, and the output means outputs an alert when the data is determined to be out of the predetermined distribution. [Supplementary Note 12] The information processing device according to Supplementary Note 1, wherein the determination means determines that the data is out of the predetermined distribution when all of the plurality of background likelihood ratios corresponding to the plurality of classification candidate classes exceed a predetermined determination threshold, and the output means outputs an alert when the data is determined to be out of the predetermined distribution. [Supplementary Note 13] The information processing device according to Supplementary Note 10, wherein the predetermined determination threshold is set according to each of the plurality of classification candidate classes. [Supplementary Note 14] The information processing device according to Supplementary Note 2, wherein the data is sequential data, the acquiring means acquires a plurality of elements included in the sequential data, and the calculating means calculates the likelihood ratio based on at least two of the plurality of elements. [Supplementary Note 15] The information processing device according to Supplementary Note 14, further comprising: a repeating means for repeatedly executing operations by the acquiring means, the calculating means, the classifying means, and the determining means until an operation completion condition is satisfied, wherein the classifying means confirms the classification of the sequential data into any of the plurality of classification candidate classes or the background class when the likelihood ratio reaches a predetermined classification threshold, and the operation completion condition is satisfied when at least one of the determining means determines that the sequential data is data outside a predetermined distribution and the classifying means confirms the classification of the sequential data.[Supplementary Note 16] The information processing device according to Supplementary Note 14, wherein at least one of the calculation of the likelihood ratio by the calculation means and the classification of the data by the classification means is performed using a sequential probability ratio test (SPRT). [Supplementary Note 17] The information processing device according to Supplementary Note 14, wherein the learning means performs learning related to the calculation of the likelihood ratio using at least one of a loss function that increases the likelihood ratio when the correct class to which the sequential data belongs is in the numerator of the likelihood ratio and decreases the likelihood ratio when the correct class to which the sequential data belongs is in the denominator of the likelihood ratio, and a log-sum-exponential loss function. [Supplementary Note 18] The information processing device according to Supplementary Note 2, wherein the classification means classifies the data by comparing the likelihood ratio calculated by the calculation means with a predetermined classification threshold. [Supplementary Note 19] An information processing method for judging data used by an information processing device comprising: an acquisition means for acquiring data; and a classification means for classifying the data into one of a plurality of classification candidate classes and a background class, the method determining whether or not the data is outside a predetermined distribution based on a background likelihood ratio between the likelihood that the data belongs to the classification candidate class and the likelihood that the data belongs to the background class, and outputting information related to the judgment. [Supplementary Note 20] A recording medium having recorded thereon a computer program for causing a computer to execute an information processing method for judging data used by an information processing device comprising: an acquisition means for acquiring data; and a classification means for classifying the data into one of a plurality of classification candidate classes and a background class, the method determining whether or not the data is outside a predetermined distribution based on a background likelihood ratio between the likelihood that the data belongs to the classification candidate class and the likelihood that the data belongs to the background class, and outputting information related to the judgment.

[0075] Although this disclosure has been described above with reference to the embodiments, this disclosure is not limited to the above-described embodiments. Various modifications that can be understood by those skilled in the art can be made to the configuration and details of this disclosure within the scope of this disclosure. Furthermore, each embodiment can be combined with other embodiments as appropriate.

[0076] 1, 2, 3, 4 Information processing device 11, 211, 311, 411 Acquisition unit 12, 212, 312, 412 Classification unit 13, 213, 413 Determination unit 14, 214 Output unit 215, 315, 415 Calculation unit 316, 416 Learning unit 417 Iteration unit

Claims

1. An information processing apparatus comprising: an acquisition means for acquiring data; a classification means for classifying the data into any one of a plurality of classification candidate classes and a background class; a determination means for determining whether the data is data outside a predetermined distribution based on a background likelihood ratio between a likelihood that the data belongs to the classification candidate class and a likelihood that the data belongs to the background class; and an output means for outputting information regarding the determination.

2. The information processing apparatus according to claim 1, further comprising a calculation means for calculating a likelihood ratio between a likelihood that the data belongs to one of two classes out of the plurality of classification candidate classes and the background class and a likelihood that the data belongs to the other.

3. The information processing apparatus according to claim 2, wherein the classification means classifies the data based on the likelihood ratio calculated by the calculation means.

4. The information processing apparatus according to claim 2, wherein the determination means makes a determination based on the background likelihood ratio calculated by the calculation means.

5. The information processing apparatus according to claim 1, wherein the determination means compares the background likelihood ratio with a predetermined determination threshold to determine whether the data is data outside the predetermined distribution.

6. The information processing apparatus according to claim 2, further comprising a learning means for performing at least one of learning regarding calculation of the likelihood ratio by the calculation means and learning regarding classification of the data by the classification means.

7. The information processing apparatus according to claim 6, wherein the learning means performs learning using learning data including classification candidate data corresponding to the plurality of classification candidate classes and background data prepared using the classification candidate data.

8. The information processing apparatus according to claim 7, wherein the background data includes at least one of data obtained by adding perturbation to the classification candidate data and data obtained by randomly mixing the classification candidate data.

9. The information processing apparatus according to claim 7, wherein the predetermined distribution is defined by a distribution of feature amounts of the learning data in a feature amount space.

10. The information processing apparatus according to claim 1, wherein the determination means determines that the data is data outside the predetermined distribution when any one of the plurality of background likelihood ratios corresponding to the plurality of classification candidate classes exceeds a predetermined determination threshold, and the output means outputs an alert when it is determined that the data is data outside the predetermined distribution.

11. When any one of the plurality of background likelihood ratios corresponding to the plurality of classification candidate classes exceeds a predetermined determination threshold, the determination means determines that the data is data outside the predetermined distribution. When it is determined that the data is data outside the predetermined distribution, the output means outputs an alert. The information processing apparatus according to claim 1.

12. When all of the plurality of background likelihood ratios corresponding to the plurality of classification candidate classes exceed a predetermined determination threshold, the determination means determines that the data is data outside the predetermined distribution. When it is determined that the data is data outside the predetermined distribution, the output means outputs an alert. The information processing apparatus according to claim 1.

13. The predetermined determination threshold is set according to each of the plurality of classification candidate classes. The information processing apparatus according to claim 10.

14. The data is time-series data. The acquisition means acquires a plurality of elements included in the time-series data. The calculation means calculates the likelihood ratio based on at least two of the plurality of elements. The information processing apparatus according to claim 2.

15. A repetition means for repeatedly executing the operations of the acquisition means, the calculation means, the classification means, and the determination means until a completion condition of the operation is satisfied. When the likelihood ratio reaches a predetermined classification threshold, the classification means determines the classification of the time-series data into any one of the plurality of classification candidate classes and the background class. The completion condition of the operation is satisfied when at least one of the cases where the determination means determines that the time-series data is data outside the predetermined distribution and the case where the classification means determines the classification of the time-series data is satisfied. The information processing apparatus according to claim 14.

16. At least one of the calculation of the likelihood ratio by the calculation means and the classification of the data by the classification means is performed using a sequential probability ratio test (SPRT). The information processing apparatus according to claim 14.

17. The learning means performs learning regarding the calculation of the likelihood ratio by using at least one of a loss function in which the likelihood ratio increases when the correct class to which the series data belongs is in the numerator of the likelihood ratio and decreases when the correct class to which the series data belongs is in the denominator of the likelihood ratio, and a log-sum-exp type loss function. The information processing apparatus according to claim 14.

18. The classification means classifies the data by comparing the likelihood ratio calculated by the calculation means with a predetermined classification threshold. The information processing apparatus according to claim 2.

19. An information processing method for determining the data used by an information processing apparatus including an acquisition means for acquiring data, and a classification means for classifying the data into any one of a plurality of classification candidate classes and a background class, the method comprising: determining whether the data is data outside a predetermined distribution based on a background likelihood ratio between the likelihood that the data belongs to the classification candidate class and the likelihood that the data belongs to the background class; and outputting information regarding the determination. The information processing method.

20. A recording medium on which a computer program for causing a computer to execute an information processing method is recorded, the information processing method including an acquisition means for acquiring data, and a classification means for classifying the data into any one of a plurality of classification candidate classes and a background class, the method comprising: determining whether the data is data outside a predetermined distribution based on a background likelihood ratio between the likelihood that the data belongs to the classification candidate class and the likelihood that the data belongs to the background class; and outputting information regarding the determination.

Images